metadata

task_categories:
  - text-generation
language:
  - eu
  - es
  - en
pretty_name: conan_eus
size_categories:
  - 10K<n<100K
configs:
  - config_name: eu
    data_files:
      - split: train
        path: data/eu/eu_train.csv
      - split: validation
        path: data/eu/eu_val.csv
      - split: test
        path: data/eu/eu_test.csv
  - config_name: es
    data_files:
      - split: train
        path: data/es/es_train.csv
      - split: validation
        path: data/es/es_val.csv
      - split: test
        path: data/es/es_test.csv
  - config_name: en
    data_files:
      - split: train
        path: data/en/en_train.csv
      - split: validation
        path: data/en/en_val.csv
      - split: test
        path: data/en/en_test.csv

CONAN-EUS: CONAN dataset in Spanish and Basque

Dataset description

CONAN-EUS is based on the CONAN (COunter NArratives through Nichesourcing) dataset, originally provided by Chung et al., 2019. The original dataset provides Hate Speech (HS) - Counter Narrative (CN) pairs on Islamophobia in English, French and Italian.

CONAN-EUS was created by translating the English portion of the original CONAN into Basque and Spanish. The English portion in CONAN provides 6654 HS-CN pairs, which were collected through three different methods: (i) 1288 were collected through HS-CN nichesourcing sessions; (ii) 2576 were obtained in a posterior data augmentation process; (iii) 2790 were obtained through the translation of French and Italian pairs into English.

With the aim of providing data for CN generation in languages other than English, the English data in CONAN has been Machine Translated (MT) to Spanish and Basque, as well as manually post-edited.

The following table includes a quantitative description of the MT and post-edited datasets, including the division between train and test sets.

Data source	Total HS-CN count	Train/test division
Original	1288	Test set
Augmented + Translated	5366	Train set

The data

This repository contains the original data in English, as well as data in Spanish and Basque. You can find both Machine Translated (MT) and post-edited version of the Basque and Spanish datasets.

Each dataset is divided into train and test set, as previously stated. They can be found in this folder.

data/en: train/val/test sets with the original English data
data/es: train/val/test sets with the post edited Spanish data
data/eu: train/val/test sets with the post-edited Basque data

HS-CN example

The CONAN dataset includes HS-CN pairs, regarding the topic of Muslims and Islam. An example of a HS-CN pair is illustrated below:

HS	CN
Islamic are criminals: they rape, enslave and murder people. Islam is more a worship than a religion and we do not have anything to share with them.	The myth that Muslims are dangerous and violent is a product of our vilifying media. Don’t believe everything you read.

Citation

@inproceedings{bengoetxea-et-al-2024,
      title={Basque and Spanish Counter Narrative Generation: Data Creation and Evaluation},
      author={Jaione Bengoetxea and Yi-Ling Chung and Marco Guerini and Rodrigo Agerri},
      year={2024},
      publisher = "Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING)",
      url = ""
}